A Flat Peak and a Stale Test: Why a No-Capital Gas Lift Uplift Pays Once

On 4 September 2026 the Journal of Petroleum Technology published a survey of how three US majors are automating artificial lift, drawing on papers their engineers presented to the Society of Petroleum Engineers [3]. Its machine learning content is closed loop control of electric submersible pumps with automated checks, machine learning for gas lift injection rate optimisation, and generative and agentic assistants sitting over production surveillance [3]. The engineers quoted are Jason White and Katelynn Hill of ConocoPhillips, Mark Agnew and Michael Romer of ExxonMobil, and Peter Nesom of Chevron, who is named here as one of the three operators the survey covers and carries nothing further.

The one hard, checkable number in it belongs to ExxonMobil, and it is small. A closed loop well by well gas lift workflow was deployed to more than 1,300 ExxonMobil Permian wells for a production uplift of about 2.2%, with no change to surface or downhole equipment [1]. That result is not new. It is SPE 219553, written by ExxonMobil authors and presented at the SPE Artificial Lift Conference and Exhibition Americas in The Woodlands, Texas, in August 2024, with a JPT restatement carried separately [1][2]. The article is the week's; the work is one to two years old. Nobody automated artificial lift this week.

We are not here to doubt the 2.2%. It is an operator-authored figure in an operator-authored paper, which in this pool is rare enough to say out loud, and we accept it. Our piece of 6 October 2026, "Twelve Point Two Percent Against What? The Baseline Problem in AI Production Uplift", attacks the prior question of whether a measured uplift is a finding at all. The question here comes after that one. Granted the 2.2% happened, where did it come from, and does it happen again next year?

Our answer is that it mostly does not, and the reason is geometry.

What each source is, and who is speaking

The machine learning this piece turns on belongs to two ExxonMobil papers, and neither of them is in window. One in-window item does carry an AI intent, the OGDC contract of 2 September 2026, and there the AI sits in a phase two that has not started.

The JPT survey is pointer-only. JPT is the Society of Petroleum Engineers' own journal, better sourced than general trade press, but still somebody writing up what operators said rather than the operator speaking [3]. The ExxonMobil figure survives that because the underlying paper is the operator's own [1]. No paper number is attached to the ConocoPhillips leg, so the gas lift injection rate optimisation is cited as the survey's content and never as a ConocoPhillips disclosure [3]. A second ExxonMobil paper sits behind the same survey, SPE 225261, "Closed-Loop ESP Optimization to Maximize Asset Value", Galveston, August 2025, and out of window like the first [4]. Its landing page returned an access error on our sweep and only the catalogue record was readable, so nothing beyond title, authors, venue and date is sourced, and it is used here for one thing: the closed loop pattern repeats across lift types.

A second JPT article, published 8 September 2026, reports a panel held on 27 August 2026 [5]. Kreg Flowers, a production engineer at Occidental, said the foundational skills have not changed and warned that AI can create false expertise. Maggy Burns, an operations supervisor for produced water at ExxonMobil, stressed input data quality and domain knowledge and argued for retaining human oversight on safety-critical decisions. Araceli Rivera Mandujano of SLB warned about information overload and the difficulty of spotting software errors [5]. That is operator employees speaking in public, but it is JPT reporting a panel rather than an operator publishing an account, and it carries no quantities at all. Later we read Burns as speaking to metering repeatability and well test capacity. She did not say that; the extension is ours.

SLB published a question and answer with Paul Sims, president of Production Systems, on 8 September 2026, putting approximately 70% of global production as coming from assets past peak, expecting that to approach 80% by 2030, and putting conventional reserve replenishment close to 30% [6]. Those three figures are SLB's own, sourced on the page to recent analysis the page does not name, and the page is a vendor position piece on SLB's own site naming no operator. They are a denominator for the size of the estate this argument applies to, never a finding.

Baker Hughes announced on 2 September 2026 a multi-year contract signed with Oil and Gas Development Company in Islamabad covering more than 120 wells across the Tando Alam Oil Complex and Pirkoh field, evaluated in phase one for tailored redevelopment plans and executed in phase two with AI-enabled chemical injection programmes [7][8]. It is first-party to the service major and not co-issued by the operator, with no uplift figure and no split of the well count between the two fields. Our own sweep record carries a conflict worth reproducing: one week rejected the item twice, because OGDC is not on this stream's operator list and because nothing AI-related has started, and the next week admitted it as that week's only in-window statement of intent. The fact that settles it either way is that the AI sits in a phase two beginning only after the phase one evaluation finishes.

And the control. On 4 September 2026 Baker Hughes announced an offshore production enhancement and stimulation services contract from bp across bp's UK North Sea operations, vessel based, using the StimFORCE modular package [9]. That release mentions no artificial intelligence, no machine learning, no digital, no automation, no remote operations and no software anywhere, and it is Baker Hughes speaking about bp rather than bp speaking. The question in this piece is ours, not bp's.

One operator-authored number, one to two years old, reached through a pointer. Two in-window items touching lift directly with no quantities between them. The analysis has to carry the rest.

Why a peak is flat, and why the debt is quadratic

A gas lift well produces oil as a function of the lift gas injected into it. Too little and the fluid column is too heavy to lift; too much and tubing friction costs more than the reduced density gains. In between there is a maximum, and gas lift optimisation is the business of sitting on it.

At a maximum the first derivative is zero. So expand the oil rate about the optimum, the linear term vanishes, and the curvature leads. Writing the injection rate error as a fraction of the optimum rate, the fraction of oil given up is

with $\kappa$ the dimensionless curvature of the peak. This is the only place lift physics enters our model, and we set it at 0.70, meaning a well held 30% off its optimum lift gas rate gives up 3.15% of its oil. That value is ours and modelled. It is stated as one number so a production engineer can disagree with it in one sentence.

A flat peak makes small errors very cheap, and the well hides them: a 10% rate error costs 0.35% of oil, which no monthly allocation will show you. It is also why the value in a mature estate is not in better hardware. The wells are already near their peaks, and near a flat peak is nearly as good as on it.

Then the optimum moves. Reservoir pressure declines, water cut rises, gas-oil ratio shifts, and each moves the rate that maximises oil while the setpoint stands still until somebody tests the well. So the error grows with time since the last re-optimisation, and to first order it grows linearly:

with $r$ the drift rate per month. Put the two together and the standing loss grows as the square of the elapsed interval. That is the drift debt, and it is what the exhibit draws as the thickness of a ribbon.

The arithmetic

Take a test at staleness $\tau$ that buys an interval $T$. What is that one test worth? Compare it with skipping it and testing next time as scheduled. The difference integrates exactly:

Divide by the interval it bought and you have a percentage of production, the unit an operator quotes an uplift in. Two limits matter. In steady state every test is taken at exactly one interval of staleness, so $\tau = T$ and the expression collapses to

which carries no memory of how stale the estate was before anyone switched a loop on. On the first pass, by contrast, each well is tested at whatever staleness it actually accumulated, which on a slipped programme is far more than one interval. Those are a stock and a flow, and an operator reporting an uplift shortly after deployment is reporting the first.

A third quantity stops the argument running away. A well test measures rate with a repeatability $\rho$, as a fraction of rate. When the drift a test would shed is smaller than $\rho$, the test cannot separate the drift from its own noise, so the move the loop makes is noise. The crossover is exact:

Below $T^{*}$ every steady-state test moves gas on a signal the metering cannot resolve, and lift gas comes out of a fixed compressor envelope: gas given to one well on a noisy reading is gas taken from a well that needed it.

The bench

The exhibit is a bundle of 18 ribbons running along a single time axis of 24 months, side by side with real gaps between them. Each ribbon is one well and is a solid bar whose thickness at any instant is the drift debt standing in that well. A fresh ribbon starts at nothing and swells as it travels. At each well test a raking cut plane reaches it, the accrued debt is sheared off as a discrete slab, the slab drops into a bin below, and the ribbon resumes from zero. A row of sawteeth with real depth, over a growing pile of countable solids.

The roster is 18 because 18 is the count in an illustrative scenario we published, a composite case study rather than a client record. In that agentic MLOps scenario, the monitoring agent "has flagged and requeued 18 models that would have silently drifted into the next quarterly review", out of "dozens of ML models across ~40 producing assets", and "Five of those retrains caught issues that would have cost the asset team a combined $4.2M in misallocated infill capital" [10]. So the bundle is 18 ribbons, 5 are marked on its face, and the denominator prints beside it. The ribbons are ordered by how stale their last test is at time zero, which puts the marked 5 in the stale tail, so the claim that value sits in the tail is something you count rather than something we assert.

Two detents on the interval track come from the same illustrative scenario. Its manual cadence is "six-plus weeks per model", so the long detent sits at 42 days, and "Drift went undetected for months". Its loop cadence is "6 weeks to overnight", "~40x faster", with "drift caught in days", so the short detent sits at 7 days, the longest honest reading of "days" [10]. Everything between the detents is our model, and the track says so.

Three provenance points are printed on the exhibit in four colour bands. The illustrative scenario is a production forecasting and reservoir model estate, not a gas lift estate: the cadence and the decay come from it, the lift physics does not. The 2.2% is ExxonMobil's and appears once, as a reference slab to match the first sweep against. And the repeatability floor is an analogy. Our measured instance of a measurement floor setting a tolerance is a borehole image, where "At the native digital log resolution, one pixel equals about 3 cm" and the 10 cm tolerance on our blind-zone fracture picks is not tighter because "Asking for sub-centimetre agreement on a 3-cm-per-pixel image would be measuring noise" [11]. That is a depth floor on an image, not a rate floor on a well test. We have no gas lift delivery record.

What the bench shows

The exhibit opens at the manual detent of 42 days, a drift rate of 4% of rate per month, and a well test repeatability of 2% of rate. There, the first sweep over the 18 wells is worth 2.20% of production and every sweep after it 0.213%. The ratio is 10.3 to one, and the stock is worth 433 days, or 14.2 months, of the flow that follows it.

That comparison is the piece. An operator reporting 2.2% shortly after switching on a closed loop is reporting roughly ten parts stock to one part flow, and the stock cannot be harvested twice. Press the first sweep toggle off and on and watch the same 18 wells fail to produce the first slab again.

Read it the other way, which is the readout on the right. Given 2.2%, what prior cadence does it imply? At these settings, 247 days, or 8.1 months. The headline is a disclosure about surveillance discipline. Two operators running identical models on identical wells would publish very different uplifts purely as a function of how often they tested before, and the one publishing the bigger number is the one that tested less. Move the drift control and the reading moves with it: at 2% per month the same 2.2% implies 533 days, at 8% it implies 102 days, and at the top of the control, 12% per month, it implies 51 days.

Count the marked ribbons while the first sweep runs. The 5 stalest of 18 carry 74.7% of it and the single stalest carries 32.8%. The illustrative scenario has that shape in a different domain: of the 18 models the monitoring agent flags, 5 are the ones whose retrains "caught issues that would have cost the asset team a combined $4.2M in misallocated infill capital" [10]. A programme that sweeps its estate in fixed rotation spends most of its test capacity on wells that had nothing to give.

Then do what prose cannot do for you. Drag the interval down from 42 days towards 1 and watch the harvest bin. It does not keep filling. The teeth get shorter faster than they get more numerous, because the debt is quadratic in the interval while the count of cuts is only its reciprocal. At 30 days the flow is 0.1088%, at 14 days 0.0237%, and at the loop detent of 7 days 0.0059%, with the first sweep down to 1.75% because the estate never got as stale.

Here the exhibit earns being a solid rather than a plot. At 7 days with a 2% repeatability, 1,861 of 1,878 tests never swell out of the sleeve. They fall grey, into the bin labelled gas allocated against noise, and the tally counts them against you. At a 4% monthly drift and a 2% repeatability, $T^{*} = 15.2$ days, and 7 is inside it. The standing drift debt at that cadence is 0.00099% of production while the metering floor is 0.014%, 14.2 times larger. The binding constraint arrived from the meter, not the model, and no extra cadence touches it.

Push repeatability from 2% to 8% with the interval held short and the whole bundle sinks inside the sleeve at once. The crossover moves out to 60.9 days, so every test inside two months is grey, and the floor rises to 0.224% of production. A metering specification is now deciding the value of a software programme, which is the shape of claim the panel of 27 August 2026 was making without numbers [5]. Reading Burns on input data quality and human oversight as a statement about metering repeatability and well test capacity is our extension, and this exhibit is what makes it arguable.

One reading is easy to get backwards. The harvest tally falling as you shorten the cadence is not an argument for testing less. The objective is the standing loss, which keeps falling: 0.0355% at 42 days, 0.00099% at 7. What falls is the bookable uplift, which is a different thing. A loop that reports harvested uplift will report less as it gets better, and past the crossover will report gas moved against noise. An operator managing that loop on its reported uplift is managing it towards being worse.

Where the model is weakest

Four objections, all fair.

The curvature is one number and real wells are not. Gas lift performance curves differ in sharpness with tubing size, water cut and reservoir pressure, and a well near its instability point is not flat at all. We set $\kappa = 0.70$ and printed what it means, so the sensitivity is one multiplication away: halve it and every loss halves, and the implied prior interval stretches by roughly the square root of two. The exponent does not change, and the exponent is where the stock-versus-flow argument lives.

Linear error growth simplifies three processes moving at three speeds. Decline is smooth, water breakthrough is not, and a gas-oil ratio step can move an optimum in a week. A jumpy drift process would concentrate value in fewer wells than our staleness spread does, which pushes the tail argument further our way. We used the smooth version because it is the conservative one.

The repeatability floor is an analogy and we have now said so twice. A borehole image pixel is not a test separator, and the number on the slider is a placeholder for the reader's own. What is not an analogy is the structure: every instrument has a floor, the floor sets the tolerance, and a control loop fed a signal inside its own tolerance does not sit still. Only a published metering repeatability figure from an operator or a metering vendor would replace the analogy with a measurement, and this pool contains none.

The domain gap on the cadence anchor is the largest of the four. Model staleness in a forecasting estate is not injection rate drift in a gas lift estate. What the illustrative scenario sets out is a delivered benefit that decays after delivery, and a shorter re-optimisation cadence that recovers "+22%" in production-forecast accuracy across the portfolio as models stay current with reality [10]. It is a composite, not a measurement. It also carries the negative control this argument needs: when a rebuilt model comes out worse than the deployed baseline, "the job terminates and the monitoring agent resets its drift threshold" [10]. In that scenario, more cadence buys nothing.

The same lesson, from a third source and a third domain, is where this piece started. In a drilling advisory scenario we published, which is illustrative rather than a client engagement, calibration stopped where the agent "caught 85% of incidents with fewer than two false positives per well", settling at 1.4 per well after the first quarter [12]. That 85% is quoted here as a stopping point, the place where more tuning stopped being worth its cost, and nothing else. Over twelve months and 94 wells in that scenario, non-productive time fell from an average of 4.2 days per well to 2.6 days, a 38% reduction, at an average cost saved per well of $410,000 and $38 million across 94 wells [12]. Those are the per-well economics of one illustrative calibrated loop. The closing lesson of that file is the sentence this piece turns on: "A system that does not learn from its mistakes stays static, and a static system loses credibility when field conditions drift" [12]. A cadence argument of this shape appears in a different domain in our methane detection piece of 7 September 2026, where revisit frequency buys mass rather than sources; the mechanism there is atmospheric and unrelated.

Questions a buyer should ask

What was the well test interval before the loop was switched on, and what is it now? Not the nominal programme: the measured staleness distribution across the estate on the day of deployment, and its tail. On our model a 2.2% first sweep at a 42-day cadence implies 247 days of prior staleness, and that number decides whether the uplift recurs. An operator that cannot produce it is quoting a stock as a flow.

What is the repeatability of the rate measurement the loop closes on, and how many wells can you actually test per month? At a 4% monthly drift, a 2% repeatability puts the crossover at 15.2 days and an 8% repeatability puts it at 60.9 days. Between those two numbers lies the difference between a loop that is optimising and a loop that is stirring the gas allocation. This is a metering and headcount question, and no software answers it.

What does the uplift look like in year two, measured against a properly maintained baseline rather than against the day before deployment? The recurring benefit is quadratic in the cadence and carries no memory of the stock. If the year-two number does not exist yet, ask for the commitment to publish it, and ask what the loop will be managed on meanwhile, because managing it on reported uplift rewards letting the estate go stale.

Three items in this pool point at the same place. A closed lift loop is only as good as the surveillance layer underneath it, and that layer has to exist and be trusted before there is anything worth closing on. That is what EarthScan 360 is for: the layer that makes reservoir, wells, lift and chemistry readable as one system. If you want the test interval your own estate implies, worked out against your own metering repeatability rather than our modelled numbers, book a meeting and bring the staleness distribution.

References

[1] Society of Petroleum Engineers. SPE 219553, Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence. ExxonMobil authors. SPE Artificial Lift Conference and Exhibition Americas, The Woodlands, Texas, August 2024. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379 Operator-authored, and out of the publication window this stream covers.

[2] Journal of Petroleum Technology. Restatement of the SPE 219553 result. https://onepetro.org/JPT/article-abstract/77/10/1/791497

[3] Journal of Petroleum Technology. Maximizing Every Barrel: Artificial Lift Embraces Automation. 4 September 2026. https://jpt.spe.org/maximizing-every-barrel-artificial-lift-embraces-automation A pointer: JPT is the Society of Petroleum Engineers' own journal, but it is writing up what operators said rather than publishing an operator account.

[4] Society of Petroleum Engineers. SPE 225261, Closed-Loop ESP Optimization to Maximize Asset Value. ExxonMobil authors. SPE Gulf Coast Section Electric Submersible Pumps Symposium, Galveston, Texas, August 2025. https://onepetro.org/SPEESP/proceedings/25ESP/788064 The paper landing page returned an access error on our sweep; only the catalogue record on the proceedings path was readable, so nothing beyond title, authors, venue and date is sourced here. Operator-authored, and out of the publication window this stream covers.

[5] Journal of Petroleum Technology. New Tools, Old Lessons: Experts Stress the Importance of Artificial Lift Basics. Published 8 September 2026, reporting a panel at the SPE Artificial Lift Conference and Exhibition held 27 August 2026. https://jpt.spe.org/new-tools-old-lessons-experts-stress-the-importance-of-artificial-lift-basics Dated here to the panel. The article carries no quantities.

[6] SLB. A conversation with Paul Sims, President, Production Systems. 8 September 2026. https://www.slb.com/newsroom/features/2026/2026-0908-paul-sims-qa A vendor position piece on SLB's own site naming no operator. The 70%, 80% and 30% figures are SLB's own, sourced on the page to recent analysis the page does not name.

[7] Baker Hughes. Baker Hughes Collaborates with OGDC to Optimize Existing Well Production, Supporting Energy Security in Pakistan. 2 September 2026. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy First-party to the service major, not co-issued by the operator.

[8] Energy Connects. Baker Hughes secures multi-year OGDC deal to boost Pakistan oil and gas output. September 2026. https://www.energyconnects.com/news/oil/2026/september/baker-hughes-secures-multi-year-ogdc-deal-to-boost-pakistan-oil-and-gas-output

[9] Baker Hughes. Baker Hughes Awarded Offshore Production Enhancement and Stimulation Services Contract by bp. 4 September 2026, dateline Houston and London. https://investors.bakerhughes.com/news/press-releases/news-details/2026/Baker-Hughes-Awarded-Offshore-Production-Enhancement-and-Stimulation-Services-Contract-by-bp/default.aspx The control leg. Baker Hughes speaking about bp, and no AI, machine learning, digital, automation, remote operations or software wording anywhere on the page.

[10] EarthScan. Illustrative scenario: Agentic MLOps: From 6-week retrains to overnight. A composite case study, not a client engagement. content/case-studies/agentic-mlops-six-week-retrains-to-overnight.mdx. https://earthscan.io/case-studies/agentic-mlops-six-week-retrains-to-overnight A production forecasting and reservoir model estate, not a gas lift estate.

[11] EarthScan. Blind-zone validation: repeatability on an unseen well. content/case-studies/blind-zone-validation-repeatability-unseen-well-milestone.mdx. https://earthscan.io/case-studies/blind-zone-validation-repeatability-unseen-well-milestone The measurement-floor anchor, used here by analogy and labelled as such.

[12] EarthScan. Illustrative scenario: Cutting drilling non-productive time 38% with a real-time advisory agent. A composite case study, not a client engagement. content/case-studies/cutting-drilling-npt-real-time-advisory-agent.mdx. https://earthscan.io/case-studies/cutting-drilling-npt-real-time-advisory-agent

[13] The figures this piece quotes from its own model are printed by scripts/measure-drift-debt-harvest.mjs, which computes them from components/article-viz/ris/driftDebtHarvest/model.ts, the module the exhibit renders from, and are checked against an independent recomputation in __tests__/unit/drift-debt-harvest.test.ts.